IAP-24-017
Harnessing Artificial Intelligence for the Automated Tracking of Icebergs through Drift and Fracture
The capabilities of artificial intelligence (AI) are accelerating in parallel with warming Polar climates. In this project we will harness the potential of AI to investigate new techniques for tracking large and tabular icebergs that calve from Arctic ice tongues. The calving of these “ice islands” is impacted by climate change, and the imposing ice features pose as potential hazards to shipping and offshore infrastructure (Fuglem & Jordaan, 2017). They also impact the local and regional marine environment through the dispersal of nutrients and freshwater (Crawford et al., 2018a; Smith, 2011).
Government and industry bodies model and monitor the drift and deterioration of ice islands to mitigate risks to stakeholders. One effective monitoring approach is the use of Earth Observation satellites (Evans et al., 2023; Koo et al., 2023; Barbat et al., 2021). However, differences in environmental conditions between the Polar Regions has challenged the transfer of solutions across hemispheres. This project will overcome this challenge, utilising a unique ice-island dataset and machine-learning approaches to develop automated detection and tracking solutions appropriate for Northern Hemisphere conditions, and specifically for tracking ice islands of north Greenland origin.
The Canadian Ice Island Drift, Deterioration and Detection (CI2D3) Database is the most comprehensive dataset of ice-island observations and information in the Canadian Arctic, and arguably, the world. Version 1 of the CI2D3 Database contains over 25,000 geospatial records of ice islands from the Petermann, Steensby, Ryder and C.H. Ostenfeld glaciers of northwest Greenland. These ice islands were manually identified and tracked through over 4,400 satellite synthetic aperture radar (SAR) scenes, these scenes primarily being acquired by the Canadian RADARSAT-2 (R2) satellite (Crawford et al., 2018b). The database is unique in its capture of lineage, which links successive observations of monitored ice islands. The size and lineage capture of the CI2D3 Database place it as the ideal training dataset for machine-learning approaches to ice-island identification and tracking as ice islands proliferate through fracture during their south-bound drift in Nares Strait, Baffin Bay and the Labrador Sea (Barbat et al., 2021).
This PhD project aims to fill the identified gap in technical capacity by using the CI2D3 Database (v1), R2 scenes, and machine-learning techniques to:
1. automatically detect and track ice islands of north Greenland origin in SAR scenes considering environmental conditions and the proliferation of ice islands through fracture processes,
2. determine the ability for technology transfer across environments and SAR sensors,
3. contribute a tool through which ice islands can be automatically detected and tracked through SAR scenes, and
4. explore complementary remote-sensing data (e.g., optical imagery, altimetry transects) to ensure detection when ice islands are obscured from identification in SAR scenes due to, for example, sea-ice backgrounds or meltwater presence.
The tool will aid end users in their efforts to provide stakeholders with information related to marine ice hazards. The tool will also be a valuable contribution to researchers looking to create or expand datasets of iceberg occurrence, with subsequent study of ice-island drift and deterioration leading to improved climatological modelling.
Potential students will be aided by a background in computer programming, and specifically machine-learning techniques. Experience with SAR data processing and analysis will be valuable but is not a requirement.
Click on an image to expand
Image Captions
A small ice island in Kane Basin, between Ellesmere Island and northern Greenland
Methodology
Based at the University of Stirling, the student will be integrated with expert support in remote sensing through their involvement with the Earth and Planetary Observation Sciences Research Group, and specifically with SAR data through the expert guidance from A. Marino and A. Crawford. Advice regarding SAR data processing and analysis will also be provided through collaboration with Prof D. Mueller (Carleton University, Canada) and the Canadian Ice Service (CIS; Environment and Climate Change Canada). The student will receive concerted training in machine-learning through B. Evans and others in the BAS AI Lab.
The R2 scenes will be used to explore techniques of detecting ice islands using co-, cross- and quad-pol SAR imagery. Using machine-learning algorithms, this will inform the development of workflows (the “tool”) through which ice islands can be automatically identified and tracked across satellite scenes. The R2 data, in combination with information contained in the CI2D3 Database v1, will be applied for training and validation purposes during the tool’s development. The workflow for that development will include these general steps:
1. extract the SAR pixel intensity (sigma nought values per available polarimetric channels), texture statistics and polarimetric observables as available for the different modes of R2 scenes;
2. establish required pre-processing steps and employ machine-learning approaches to develop a method for detection of ice islands of north Greenland origin in R2 SAR scenes, utilising the georeferenced polygons within version 1 of the CI2D3 Database and exploring the use of self-supervised machine-learning processes;
3. explore shape-based approaches for the tracking of ice islands through drift and fracture (Barbat et al., 2021), and
4. finalise an automated ice-island detection and tracking tool.
Given evolving remote-sensing technology, the student will also test how well the developed tool transfers to other sensors (e.g., RADARSAT Constellation Mission or Sentinel-1). In addition, the student will explore how the incorporation of remotely-sensed data, complementary to SAR, can benefit ice-island detection and tracking efforts.
The student will be encouraged and supported in the publication of their work as peer-reviewed journal articles. The student will also present their work at local (e.g. Biological and Environmental Science’s student symposium), national (e.g. U.K. Arctic Science Conference) and international conferences (e.g. European Geophysical Union’s General Assembly). There is also potential for the student to present to the International Ice Charting Working Group.
Project Timeline
Year 1
Research design; Training activities; Processing of SAR scenes; Exploration of machine-learning approaches; Writing literature review
Year 2
Workflow for ice-island detection in SAR; Analysis and synthesis of results; Conference presentations; Writing of first data chapter
Year 3
Programme for ice-island tracking; Analysis and synthesis of results; Conference presentations; Meeting with End Users and Collaborators; Writing of second data chapter; Publication of first data chapter
Year 3.5
Testing of developed programmes on SAR data acquired by additional sensors; Presentation to End Users and Collaborators; Publication of second data chapter; Completion of third data chapter of thesis and writing of remaining sections
Training
& Skills
Training for the student will be organised in topics such as the processing and analysis of remote-sensing data (e.g. ESA’s Advanced Training Course on Radar Polarimetry), software engineering (e.g. The Alan Turing Institute), and/or machine-learning techniques (in-house with BAS). The student will develop skills in presenting their research to diverse audiences and in academic and non-academic writing.
References & further reading
Barbat, M, et al. (2021). Automated iceberg tracking with a machine learning approach applied to SAR imagery: A Weddell sea case study. ISPRS Journal of Photogrammetry and Remote Sensing, 172, 189–206.
Crawford, A, et al. (2018a). The aftermath of Petermann Glacier calving events (2008–2012): Ice island size distributions and meltwater dispersal. Journal of Geophysical Research: Oceans, 123(12), 8812–8827.
Crawford, A, et al. (2018b). The Canadian Ice Island Drift, Deterioration and Detection (CI2D3) Database. Journal of Glaciology, 64(245), 517–521.
Evans, B, et al. (2023). Unsupervised machine learning detection of iceberg populations within sea ice from dual-polarisation SAR imagery. Remote Sensing of Environment, 297, 113780.
Fuglem, M, Jordaan, I (2017). Risk Analysis and Hazards of Ice Islands. In L. Copland & D. Mueller (Eds.), Arctic Ice Shelves and Ice Islands (pp. 395–415). Dordrecht: Springer Netherlands.
Koo, Y, et al. (2023). Automated detection and tracking of medium-large icebergs from Sentinel-1 imagery using Google Earth Engine. Remote Sensing of Environment, 296, 113731.
Smith, KL (2011). Free-drifting icebergs in the Southern Ocean: An overview. Deep Sea Research Part II: Topical Studies in Oceanography, 58(11–12), 1277–1284.
